A sparsity-triggered local fluctuation distribution entropy wind turbine blade fault diagnosis method

Abstract

This invention relates to the field of wind turbine blade fault detection technology, and discloses a sparsity-triggered local fluctuation spread entropy wind turbine blade fault diagnosis method. The method includes: acquiring vibration signals of the wind turbine blade and preprocessing the vibration signals; calculating the temporal sparsity of the preprocessed vibration signals and comparing the temporal sparsity with a first dynamic threshold; if the temporal sparsity is greater than or equal to the first dynamic threshold, locating the impact center moment in the vibration signal; extracting a local signal segment centered on the impact center moment; calculating the local fluctuation spread entropy of the local signal segment; comparing the local fluctuation spread entropy with a second dynamic threshold; if the local fluctuation spread entropy is less than the second dynamic threshold, determining that the wind turbine blade has a fault, and outputting a fault flag. This invention can effectively distinguish between real damage and wind turbulence noise under single-sensor and no prior model conditions, improving diagnostic accuracy.

Description

Technical Field

[0001] This invention relates to the field of wind turbine blade fault detection technology, and more particularly to a method for diagnosing wind turbine blade faults based on sparsity-triggered local fluctuation spread entropy. Background Technology

[0002] Wind turbine blades operate under conditions of strong winds, rain erosion, dust storms, and extreme weather, making them susceptible to structural damage such as microcracks, debonding, and delamination. In the early stages, this type of damage only triggers a non-periodic, low-amplitude, millisecond-level transient impact response. This impact is masked by strong background wind turbulence noise and lacks a fixed occurrence time and theoretical characteristic frequency model, rendering traditional diagnostic methods based on frequency domain characteristics (such as envelope spectrum peak values ​​and characteristic frequencies) ineffective.

[0003] Existing technologies mainly rely on two approaches: one is to extract the natural frequency changes of the blades through a multi-point sensor array, but this method is difficult to capture stably under random wind load excitation, and the hardware cost is high and the engineering implementation is difficult; the other is to use multi-band energy ratio or envelope spectrum analysis, but such methods are sensitive to wind speed disturbances, have a high false alarm rate, and are difficult to meet the stringent requirements of wind power operation and maintenance for diagnostic reliability.

[0004] Crucially, if the wave spread entropy (FDE) is calculated for the entire vibration signal, the impact characteristics will be diluted by the noise in the non-impact segment, causing the criterion to fail; while if only the time-domain sparsity is relied upon, it is impossible to distinguish between real damage impact and random pulse.

[0005] Therefore, there is an urgent need for a lightweight diagnostic method with a causal hierarchy that can effectively distinguish between real damage and wind turbulence noise under single-sensor and no prior model conditions. Summary of the Invention

[0006] The purpose of this invention is to propose a sparsity-triggered local fluctuation distribution entropy wind turbine blade fault diagnosis method, which solves the technical problem that the existing diagnosis methods are not accurate enough.

[0007] Specifically, the present invention provides a method for diagnosing wind turbine blade faults based on sparsity-triggered local fluctuation distribution entropy, comprising the following steps:

[0008] S1. Collect the vibration signal of the wind turbine blades and preprocess the vibration signal to obtain the preprocessed vibration signal;

[0009] S2. Calculate the time-domain sparsity of the preprocessed vibration signal and compare the time-domain sparsity with the first dynamic threshold.

[0010] S3. If the time-domain sparsity is greater than or equal to the first dynamic threshold, then locate the impact center time in the vibration signal.

[0011] S4. Using the moment of impact center as the center, extract a local signal segment;

[0012] S5. Calculate the local fluctuation spread entropy of the local signal segment;

[0013] S6. Compare the local fluctuation spread entropy with the second dynamic threshold. If the local fluctuation spread entropy is less than the second dynamic threshold, it is determined that the wind turbine blade has a fault, and a fault flag is output.

[0014] The beneficial effects provided by this invention are: it can effectively distinguish between real damage and wind turbulence noise under the condition of a single sensor and no prior model, thereby improving diagnostic accuracy. Attached Figure Description

[0015] Figure 1 This is a simplified flowchart of the method of the present invention. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0017] Before formally describing the present invention, a general description of the solution of the present invention will be given first to facilitate understanding.

[0018] Please refer to Figure 1 The present invention provides a method for diagnosing wind turbine blade faults based on sparsity-triggered local fluctuation scattering entropy, comprising the following steps:

[0019] S1. Collect the vibration signal of the wind turbine blades and preprocess the vibration signal to obtain the preprocessed vibration signal;

[0020] It should be noted that step S1 includes:

[0021] S11. Install a biaxial piezoelectric accelerometer at 1 / 3 of the length of the wind turbine blade from the root blade to synchronously collect vibration acceleration signals in the waving direction and the swinging direction.

[0022] S12. Sample the vibration acceleration signal using a preset sampling frequency and analysis window length;

[0023] S13. Perform zero-mean and low-pass filtering on the sampled signal to obtain the preprocessed vibration signal.

[0024] As one embodiment, the sensor configuration of the present invention is as follows: a single-axis accelerometer is installed at the root of a single blade (0.1L from the blade root) to collect vibration acceleration signals in the waving direction. ;in, This is the discrete-time index (sampling point number). Units are ;

[0025] Sampling parameters: Sampling frequency =2kHz, analysis window length =10,000 (5 seconds); where, Sampling frequency (unit: Hz) This represents the total number of sampling points included in a single analysis window;

[0026] Preprocessing: Zero-mean normalization + 8th-order Butterworth low-pass filter (cutoff frequency) =500Hz), resulting in the processed signal. .in, This is the cutoff frequency of the low-pass filter (unit: Hz). This is the preprocessed discrete vibration signal.

[0027] S2. Calculate the time-domain sparsity of the preprocessed vibration signal and compare the time-domain sparsity with the first dynamic threshold.

[0028] It should be noted that transient impacts caused by wind turbine blade damage have the physical characteristic of highly concentrated energy, meaning that a large amount of energy is released in a very short time, causing abrupt amplitude changes in the vibration signal at a few sampling points, while most other sampling points are close to zero. This "pulse concentration" phenomenon makes the signal highly sparsity in the time domain, while the background wind turbulence noise exhibits a dispersed Gaussian distribution. Therefore, this invention introduces time-domain sparsity as a sensitive criterion for the existence of impacts, effectively identifying damage impacts by quantifying the degree of signal energy concentration.

[0029] The formula for calculating temporal sparsity in this invention is as follows:

[0030]

[0031] in, S For time-domain sparsity, N This represents the total number of sampling points included in a single analysis window during sampling. x ( n ) represents the pre-processed vibration signal.

[0032] In this invention, if the time-domain sparsity is less than the first dynamic threshold, the wind turbine blades are directly determined to be normal, and the calculations of S3 to S6 are skipped.

[0033] It should be noted that the first dynamic threshold in step S2 is a dynamic threshold based on impact-background separation and signal-to-noise ratio adaptation, and its setting method includes the following steps:

[0034] S21. Estimate the background energy level of the signal within the entire analysis window. ;

[0035] S22. Identify candidate impulse segments in the signal and calculate their total energy. ;

[0036] S23. Calculate the first dynamic threshold. :

[0037] in, The sparsity of the background signal. The first proportionality coefficient, A constant to prevent division by zero.

[0038] In this invention, the sparsity threshold should be able to distinguish between "significant pulse concentrations" and "general noise fluctuations". Therefore, a relative sparsity increment based on background noise energy normalization is introduced as a criterion.

[0039] As one embodiment, the formula and steps for setting the dynamic threshold are as follows:

[0040] Background noise level estimation: Calculate the preprocessed signal within the entire analysis window. The root mean square (RMS) value is used as an estimate of the background vibration energy level. .

[0041]

[0042] Candidate impulse extraction and energy calculation: Impulse screening of the signal based on statistical principles. First, the preprocessed signal is calculated. sample mean and sample standard deviation The impact discrimination threshold is defined as: in =3 (corresponding to a 99.7% confidence interval). Find all conditions that satisfy this condition. sampling points , each This is considered a candidate point for the impact center. Impact segment width determination: Based on the physical characteristics of wind turbine blade damage impact, the impact energy is mainly concentrated within a 2.5ms time window before and after the impact peak point. Fatigue tests on 1.5MW wind turbine blades have shown that over 95% of the impact energy is concentrated within this width. Therefore, the section with... Center, width A signal segment of 5ms (corresponding to 10 sampling points, sampling frequency 2kHz) is designated as the "candidate impulse segment". : Index value of the candidate point of impact center (sampling point number). Impact segment width, in milliseconds (ms). Calculation of total candidate impact energy: For each candidate impact segment... (in and , , =10,000 is the total signal length, and its total energy is calculated as follows:

[0043]

[0044] If multiple candidate impact segments exist, the total candidate impact energy is:

[0045]

[0046] in This represents the number of candidate impact segments.

[0047] Background energy estimation: Let... The set of indices for the background segment, i.e. Then the background signal energy is:

[0048]

[0049] in Represents a set The number of elements.

[0050] The first dynamic threshold is calculated as follows:

[0051]

[0052] The formula is explained in detail as follows: The baseline sparsity of the background signal. It can be obtained by calculating the sparsity of the signal segment after excluding candidate impact segments. Essentially, it is a proxy metric for the time-domain signal-to-noise ratio (SNR). It quantifies the salience of a candidate impulse relative to the background. : A scaling factor used to map the signal-to-noise ratio to the increment of sparsity, which can be calibrated using historical data or simulation. A very small number to prevent the denominator from being zero.

[0053] The inventiveness of the first dynamic threshold of this invention lies in the threshold... It is not a fixed value, but rather positively correlated with the signal-to-noise ratio (SNR) of "suspicious events" in the current signal. When there is a strong impact, the SNR is high, and the threshold automatically increases, ensuring that only the most significant impact triggers a precise judgment, reducing false triggering of minor fluctuations in a noisy environment. When the environment is stable, the SNR is low, the threshold decreases, thereby improving the detection sensitivity to weak impacts.

[0054] S3. If the time-domain sparsity is greater than or equal to the first dynamic threshold, then locate the impact center time in the vibration signal.

[0055] It should be noted that step S3 specifically includes:

[0056] S31. Set the sliding window length and sliding step size;

[0057] S32. Using the sliding window, slide it over the preprocessed vibration signal and calculate the local sparsity of each window;

[0058] S33. Select the window center time corresponding to the maximum local sparsity as the impact center time.

[0059] As one embodiment, sparsity only reflects the presence of impact across the entire window and cannot determine the moment of impact. To accurately locate the impact center, this invention uses a sliding window to calculate local sparsity:

[0060] Local sparsity calculation: Let the sliding window length be... =100 (50ms), step size =10 (5ms);

[0061] For each window Calculate local sparsity:

[0062]

[0063] Impact center location: Take The maximum value corresponds to the center time of the window. As the moment of impact.

[0064] S4. Using the moment of impact center as the center, extract a local signal segment;

[0065] S5. Calculate the local fluctuation spread entropy of the local signal segment;

[0066] It should be noted that step S5 specifically includes:

[0067] S51. Map the local signal segment into a symbol sequence;

[0068] It should be noted that the generation of the sequence in step S51 includes:

[0069] S511. Map the local signal segment to the [0,1] interval using the standard normal cumulative distribution function;

[0070] S512. Linearly quantize the mapped values ​​into 7 types of integer symbols to obtain a symbol sequence. ∈{1,2,…,7}; where the standard normal cumulative distribution function is used to balance the sensitivity of different amplitude ranges, and linear quantization is used to discretize continuous amplitudes, making the jump characteristics of the shock easier to identify.

[0071] S52. Calculate the difference between adjacent symbols in the symbol sequence and count the probability of occurrence of each difference pattern;

[0072] S53. Calculate the local fluctuation spread entropy based on the probability distribution of the difference pattern.

[0073] As one example, the impact waveform caused by wind turbine blade damage has a deterministic structure of "oscillatory decay" (such as a decaying oscillation that rises first and then falls), and the amplitude changes between adjacent sampling points show a regular trend; while the amplitude changes of random noise such as wind turbulence are irregular, and the differences between adjacent points are randomly distributed. This difference makes the dynamic fluctuation pattern of the impact signal more "regular" and less complex than that of noise. Therefore, this invention introduces the wave dispersion entropy (… By statistically analyzing the distribution complexity of the difference between adjacent symbols, the structural regularity of the impact is quantified, effectively distinguishing between real faults and random interference.

[0074] In this invention, the total number of difference patterns is determined by the formula. Decision, among which The number of symbol categories is (7). Substituting the embedding dimension (2), we get =13. These 13 patterns correspond to all possible differences between adjacent points in the symbol sequence, reflecting the dynamic fluctuation characteristics of the signal:

[0075] Difference The value range is [-6, +6], with a total of 13 possibilities, each corresponding to a dynamic fluctuation pattern, as shown in the table below:

[0076] Table 1 Physical meaning of the difference pattern

[0077]

[0078] Blade damage impact is a damped oscillating waveform, and its symbol sequence exhibits regular fluctuations of "rising first and then falling" or "falling first and then rising", resulting in the distribution of differences between adjacent symbols being concentrated in a few patterns; while the difference distribution of wind turbulence noise is uniform and highly complex.

[0079] Therefore, this invention calculates FDE only within the impact neighborhood to avoid noise interference in the non-impact segment.

[0080] The local signal is extracted as follows: Extracted by... Center, width =5ms (10 sampling points) signal segment ;

[0081] The symbol sequence generation process is as follows: Mapping to the [0,1] interval using the normal cumulative distribution function, and then linearly transforming it into a sequence of integer symbols. ∈{1,2,…,7} ;

[0082] The difference pattern is calculated as follows: calculate the difference between adjacent symbols. There are 13 possible values ​​(from -6 to +6);

[0083] Statistical analysis of the probability of 13 difference patterns (k=1,2,…,13);

[0084] The local FDE calculation is as follows: ;in, Local fluctuation spread entropy (unit: nat);

[0085] Overall, the stronger the regularity of the impact, the more concentrated the distribution of the difference. The smaller the value (normal FDE ≈ 2.45, fault FDE < 2.10).

[0086] S6. Compare the local fluctuation spread entropy with the second dynamic threshold. If the local fluctuation spread entropy is less than the second dynamic threshold, it is determined that the wind turbine blade has a fault, and a fault flag is output.

[0087] It should be noted that the second dynamic threshold calculation process in step S6 is as follows:

[0088] S61. Calculate the baseline value of the wave dispersion entropy of the background signal segment. ;

[0089] S62, Calculate the second dynamic threshold. :

[0090]

[0091] in, The minimum fluctuation spread entropy of the preset ideal rule signal, The regularity confidence factor, calculated based on the current signal-to-noise ratio of the impact segment, is dynamically calculated according to other characteristics of the impact segment as follows:

[0092]

[0093] in This is the signal-to-noise ratio estimate for that impact segment. It is an adjustable parameter.

[0094] Specifically, this application argues that the FDE threshold should be able to determine whether the impact waveform is sufficiently regular. We propose a threshold based on the relative difference in FDE between the impact segment and the background segment.

[0095] The specific formula and steps for determining the second dynamic threshold are as follows:

[0096] Background complexity benchmark The wave propagation entropy is calculated from the background signal segment (excluding all candidate impact segments) and used as a benchmark for background complexity.

[0097] Impact segment regularity measure: For an impact segment triggered and located by the first stage, calculate its local wave spread entropy. .

[0098] Dynamic second threshold calculation:

[0099]

[0100] The formula is explained as follows:

[0101] The FDE value of the background noise represents the most chaotic state.

[0102] FDE is a theoretically or empirically "most regular" value, representing the FDE corresponding to an ideal, completely deterministic damped oscillation signal. It can be a calibration constant or calculated by generating a signal through a simple damped oscillation model.

[0103] This represents the possible range of signal variation from the most chaotic to the most regular under the current operating conditions.

[0104] A regular confidence factor between 0 and 1. It can be dynamically calculated based on other characteristics of the impact segment, such as:

[0105]

[0106] in This is the signal-to-noise ratio estimate for that impact segment. It is an adjustable parameter.

[0107] In this invention, the second dynamic threshold It is related to the complexity of the background and the quality of the impact itself (signal-to-noise ratio).

[0108] When the impulse signal-to-noise ratio is very high ( ), threshold Approaching This means that we require the impact waveform to be very regular, close to the ideal fault model.

[0109] When the impact signal-to-noise ratio is low ( ), threshold Approaching This means we have relaxed the requirements for regularity, because noise pollution can make even regular impact waveforms complex. At this point, although the FDE threshold has been relaxed, its low signal-to-noise ratio reduces its likelihood of passing the first level of sparsity screening, and the system as a whole remains conservative.

[0110] In summary, the effects and advantages of this invention are as follows:

[0111] The first-ever “sparseness-triggered local FDE” hierarchical diagnostic architecture uses time-domain sparsity as a lightweight trigger to calculate local wave spread entropy (FDE) only in the impact candidate region. The two form a causal dependency rather than a simple parallel relationship, effectively avoiding interference from non-impact segment noise in the full signal FDE and significantly improving the sensitivity of the criterion to real faults.

[0112] Construct a two-stage diagnostic logic of "coarse screening - fine judgment" to balance efficiency and accuracy: The first stage (coarse screening) quickly determines whether there is a pulse concentration phenomenon by using time domain sparsity. If not, skip the FDE calculation directly to save more than 90% of computing power.

[0113] The second stage (precise judgment) calculates FDE only within the impact neighborhood indicated by sparsity (e.g., ±2.5ms) to accurately capture the structural regularity of the "oscillation decay" of the impact and effectively suppress false alarms of random pulses caused by wind turbulence.

[0114] Achieving fault diagnosis with a single sensor, low computational complexity, and high engineering applicability: requiring only one blade root accelerometer, the algorithm does not require frequency domain transformation, dictionary learning, or multi-point synchronization. It has been verified in the fatigue test of a 1.5MW wind turbine blade for its sensitivity to early cracks, with a measured false alarm rate of <2%, demonstrating strong engineering applicability.

[0115] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for diagnosing wind turbine blade faults based on sparsity-triggered local fluctuation distribution entropy, characterized in that: Includes the following steps: S1. Collect the vibration signal of the wind turbine blades and preprocess the vibration signal to obtain the preprocessed vibration signal; S2. Calculate the time-domain sparsity of the preprocessed vibration signal and compare the time-domain sparsity with the first dynamic threshold. S3. If the time-domain sparsity is greater than or equal to the first dynamic threshold, then locate the impact center time in the vibration signal. S4. Using the moment of impact center as the center, extract a local signal segment; S5. Calculate the local fluctuation distribution entropy of the local signal segment; S6. Compare the local fluctuation spread entropy with the second dynamic threshold. If the local fluctuation spread entropy is less than the second dynamic threshold, it is determined that the wind turbine blade has a fault, and a fault flag is output. In step S2, the formula for calculating temporal sparsity is as follows: in, S For time-domain sparsity, N This represents the total number of sampling points included in a single analysis window during sampling. x ( n () indicates the pre-processed vibration signal; In step S2, the first dynamic threshold is a dynamic threshold based on impact-background separation and signal-to-noise ratio adaptation, and its setting method includes the following steps: S21. Estimate the background energy level of the signal within the entire analysis window. ; S22. Identify candidate impulse segments in the signal and calculate their total energy. ; S23. Calculate the first dynamic threshold. : in, The sparsity of the background signal. The first proportionality coefficient, To prevent division by zero constants; In step S22, the specific steps for identifying candidate impact segments are as follows: calculate the mean of signal samples. with sample standard deviation The impact discrimination threshold is defined as follows: ,in =3, corresponding to a 99.7% confidence interval; find all conditions that satisfy this condition. sampling points and will be Center, width =5ms, corresponding to 10 sampling points, and the signal segment with a sampling frequency of 2kHz is defined as the candidate impulse segment; The calculation process for the second dynamic threshold in step S6 is as follows: S61. Based on historical data of normal blades under the same operating conditions, calculate the baseline value of the wave dispersion entropy of the background signal segment. ; S62, Calculate the second dynamic threshold. : in, The minimum fluctuation spread entropy of the preset ideal rule signal, The regularity confidence factor, calculated based on the current signal-to-noise ratio of the impact segment, is dynamically calculated according to other characteristics of the impact segment as follows: in This is the signal-to-noise ratio estimate for that impact segment. It is an adjustable parameter.

2. The method for diagnosing wind turbine blade faults based on sparsity-triggered local fluctuation distribution entropy as described in claim 1, characterized in that: Step S1 includes: S11. Install a single-axis or dual-axis piezoelectric accelerometer at 1 / 3 of the length of the wind turbine blade from the root blade to collect vibration acceleration signals in the waving direction and the swinging direction. S12. Sample the vibration acceleration signal using a preset sampling frequency and analysis window length; S13. The collected vibration acceleration signal is subjected to zero-mean normalization and 8th-order Butterworth low-pass filtering to obtain the preprocessed vibration signal.

3. The method for diagnosing wind turbine blade faults based on sparsity-triggered local fluctuation distribution entropy as described in claim 1, characterized in that: Step S3 specifically includes: S31. Set the sliding window length and sliding step size; S32. Using the sliding window, slide it over the preprocessed vibration signal and calculate the local sparsity of each window; S33. Select the window center time corresponding to the maximum local sparsity as the impact center time.

4. The method for diagnosing wind turbine blade faults based on sparsity-triggered local fluctuation distribution entropy as described in claim 1, characterized in that: Step S5 specifically includes: S51. Map the local signal segment into a symbol sequence; S52. Calculate the difference between adjacent symbols in the symbol sequence and count the probability of occurrence of each difference pattern; S53. Calculate the local fluctuation spread entropy based on the probability distribution of the difference pattern.

5. The method for diagnosing wind turbine blade faults based on sparsity-triggered local fluctuation distribution entropy as described in claim 4, characterized in that: The generation of the sequence in step S51 includes: S511. Map the local signal segment to the [0,1] interval using the standard normal cumulative distribution function; S512. Linearly quantize the mapped values ​​into 7 types of integer symbols to obtain a symbol sequence. Among them, the standard normal cumulative distribution function is used to balance the sensitivity of different amplitude ranges, and linear quantization is used to discretize continuous amplitudes, making the jump characteristics of the shock easier to identify.

Images